Skip to content

Tier 2: Build the Sharpen generator path (blind self-critique) #92

Description

@NickCrew

Summary

Build the Sharpen generator path — the "blind self-critique" generation mode
that is the quality engine in the AI-action-verbs design. This is Tier 2 work
from docs/development/design/ai-action-verbs-and-sharpen-mechanic.md
(Status: Proposed, 2026-06-21).

Why

The product philosophy is correction > creation. A model shown its own draft
trends toward consistency and self-defense ("this is fine, I wrote it"). The same
model shown the identical content with the authorship tag stripped, framed as
"critique and improve this candidate's draft," switches into critical evaluation
and finds the weaknesses. Sharpen operationalizes that as a prompt wrapper. It is
the brand promise made literal — tagline "See yourself, sharpened."

This is distinct from the two modes we run today:

  • Fresh generation (mode 2, today's "Regenerate"): no prior draft, re-derives
    everything, discards good structure — and is the most inference-expensive mode.
  • Self-aware incremental (mode 1): prior draft framed as "you wrote this,
    update it"
    — anchors and entrenches.

Sharpen is mode 3: prior draft in context, authorship hidden, ask for critique.

Scope

  • A generator wrapper that feeds the prior artifact draft to the existing
    generator framed as external work to critique — never "you wrote this."
  • Per-surface opt-in (the design is explicit that the user does not pick the
    mechanic at generation time; the system selects the verb per surface + state).
  • Prime candidates: the staleness banners that are Refresh-intent but
    Regenerate-behavior today
    (cover letter, prep) — see the design's Implementation
    section.

Dependency / sequencing

The generator path itself can be prototyped independently. But making Sharpen
the default, run-confidently verb
depends on rollback existing (the data-safety
axis) — otherwise a Sharpen that regresses output is unrecoverable. See the
version-history/rollback issue; the design's "quality vs safety are orthogonal"
section is the rationale. Land rollback before flipping Sharpen on as default.

Out of scope

Acceptance criteria

  • A Sharpen generator wrapper exists that submits the prior draft as
    external/unattributed work to critique, with no "you authored this" framing.
  • At least one real surface (cover-letter or prep staleness path) runs Sharpen
    instead of from-scratch regeneration.
  • The verb the user sees on that surface reads honestly as its behavior.
  • Tests cover the prompt framing (authorship stripped) and the surface wiring.

Design: docs/development/design/ai-action-verbs-and-sharpen-mechanic.md
Related: #31 (Tier 1 standardization)

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    • Status
      Blocked

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions